First, we plot the presence/absence data from the Nishikawa dataset for slender tuna.
No positive sampling points.
No positive sampling points.
Few sampling points off the east coast of Australia.
Then, we build the basic slender tuna model with all the predictors included. Note that the environmental predictors are mean values over 1956-1981.
## [1] "training AUC: 0.9919"
## [1] "testing AUC: 0.9856"
Then, we extrapolate for the rest of \(40^{\circ}N\)-\(40^{\circ}S\) and present seasonal distribution maps. The distribution maps are shown side-by-side with the Nishikawa maps.
Generally speaking, even with the \(0-360^{\circ}\) longitude, there are still visible artefacts for the slender tuna’ model.
## [1] "training AUC: 0.9941"
## [1] "testing AUC: 0.9865"
Again, each seasonal distribution map is shown side-by-side with its corresponding Nishikawa seasonal chart.
The artefacts are gone when using Model 2.
For this section, we use Model 1 (full model). We use the same \(10 \times 10\) grid.
Similar to the yellowfin tuna, for each season, we associate the \(10 \times 10\) grid with the \(1 \times 1\) grid cells. Then, we limit the area to \(10 \times 10\) grid cells with sampling points.
We only leave \(10 \times 10\) grid cells that have sampling points within a certain % area threshold. We first do this for a more conservative 25% threshold.
We also do it for a more liberal 10% threshold.
Then, we replicate this across the 3 other seasons…